Model Identifier: Quin-Beta-21.ab1
Status: EXPERIMENTAL (v0.2.1)
Developed by: Dr. ChamYoung
Researche Paper Dynamics By : chethanyadav456
Model Training by: Seo Park Jun
Primary Purpose: Architectural validation and optimization testing
Parent Architecture: Quin-91 (910B)
Quin Beta 21.ab1 (8B) is a pre-production internal research artifact, developed as a public testbed for experimental architectural hypotheses and optimization strategies.
It is not a general-purpose foundational model. The model serves as a compact 8B-parameter prototype designed to validate two novel components:
- Resonant Gating Units (RGUs) — Frequency-modulated activation and selective neuron gating.
- Causal State Tensors (CSTs) — Predictive vector-field–based context management.
This version is intended for academic analysis and architectural ablation, not for deployment or production inference.
The Quin architecture diverges from standard transformers by treating latent space as a predictive vector field, rather than a static memory structure. This approach seeks to optimize representational efficiency and contextual foresight.
| Component | Description |
|---|---|
| RGU (Resonant Gating Unit) | Frequency-based dynamic gating replacing standard FFN layers. |
| CST (Causal State Tensor) | Fourth-order tensor predicting future state transitions. |
| APP (Asymmetric Parameter Projection) | Downscaled parameter projection from Quin-91’s 910B latent map. |
The RGU activation function φ(x) is governed by a resonance function R, modulating input x via a learned frequency matrix Ω and a gate G:
[ ϕ(x) = G(x, W_g) · tanh(R(x, Ω) ⊙ (xW_v)) ]
where resonance R(x, Ω) is defined as:
[ R(x, Ω) = cos^2(proj(x, W_k) · Ω) ]
This structure enforces resonance-driven activation pathways, forming the theoretical core of Quin-21 Beta.
| Property | Specification |
|---|---|
| Model Type | quin (Custom Architecture) |
| Variant | 21.ab1 (Asymmetric Beta 1) |
| Parameter Count | ~8.1B (Projected from Quin-91) |
| Parent Model | Quin-91 (910B) |
| Context Window | 8192 Tokens (Truncated) |
| Gating Mechanism | Resonant Gating Unit (RGU) |
| State Management | Causal State Tensor (T_cs) |
| Innovation | Description |
|---|---|
| Asymmetric Parameter Projection (APP) | Projects a high-fidelity subspace from Quin-91’s 910B latent map, preserving conceptual gradients and scaling effects. |
| Resonant Gating Units (RGU) | Replaces FFN/MoE layers, activating only frequency-aligned neurons, reducing FLOPs per token. |
| Causal State Tensors (CST) | Eliminates KV cache. Predicts context trajectory via a 4th-order causal tensor field. |
| Category | Notes |
|---|---|
| Objective | Internal testing of architectural optimization by Dr. ChamYoung and Seo Park Jun. |
| Stability | Unstable – not tuned for production. |
| Safety | No instruction tuning or alignment applied. |
| Purpose | Architectural transparency and academic evaluation only. |
| Usage | Do not deploy for real-world or production tasks. |
Direct integration into transformers is not supported.
Model loading requires a custom QuinLoader class capable of initializing CST and RGU components.
# WARNING: Pseudocode. This loader is not part of the public library.
from quin_arch_loader import QuinLoader, QuinConfig
config = QuinConfig(
model_type="quin",
variant="21.ab1",
n_layers=32,
rgu_expansion_factor=2.75,
cst_dimensions=1024,
projection_parent_hash="quin-91/proj-map-v2.bin"
)
model = QuinLoader.from_pretrained(
"drchamyoung/Quin-Beta-21.ab1",
config=config,
device_map="auto",
precision="bf16"
)
# model.get_architecture_report()
# > [Quin-21.ab1]: 8.1B Params. 32 RGU layers. CST dim 1024.| Resource | URL |
|---|---|
| Primary Repo | https://github.com/InboraStudio |
| Contributor | https://github.com/SeoParkjun |
| Discord Research Server | https://discord.gg/pVA8NZ5mru |
| Datasets | https://huggingface.co/datasets/DrChamyoung/Quinbeta5.2DataSets |
Retrieve Rows:
curl -X GET "https://datasets-server.huggingface.co/rows?dataset=DrChamyoung%2FQuinbeta5.2DataSets&config=default&split=train&offset=0&length=100"List Splits:
curl -X GET "https://datasets-server.huggingface.co/splits?dataset=DrChamyoung%2FQuinbeta5.2DataSets"List Parquet Files:
curl -X GET "https://huggingface.co/api/datasets/DrChamyoung/Quinbeta5.2DataSets/parquet/default/train"If referencing this artifact in academic or technical research:
Dr. ChamYoung, Seo Park Jun. Quin Beta 21.ab1 (8B) — Experimental Architectural Model. Inbora Studio Research Division, 2025.
This model and documentation are released for non-commercial academic research only.
All intellectual property rights reserved to Inbora Studio Research Group.
- Architectural Design: Dr. Chamyoung AKA Alok
- Optimization & Validation: Seo Park Jun
- Infrastructure: Inbora Studio Labs
- Dataset Preparation: DrChamyoung AKA Alok Research Division